1/*
2 * Equalizer Load-balancer for serial network interfaces.
3 *
4 * (c) Copyright 1995 Simon "Guru Aleph-Null" Janes
5 * NCM: Network and Communications Management, Inc.
6 *
7 * (c) Copyright 2002 David S. Miller (davem@redhat.com)
8 *
9 *	This software may be used and distributed according to the terms
10 *	of the GNU General Public License, incorporated herein by reference.
11 *
12 * The author may be reached as simon@ncm.com, or C/O
13 *    NCM
14 *    Attn: Simon Janes
15 *    6803 Whittier Ave
16 *    McLean VA 22101
17 *    Phone: 1-703-847-0040 ext 103
18 */
19
20/*
21 * Sources:
22 *   skeleton.c by Donald Becker.
23 * Inspirations:
24 *   The Harried and Overworked Alan Cox
25 * Conspiracies:
26 *   The Alan Cox and Mike McLagan plot to get someone else to do the code,
27 *   which turned out to be me.
28 */
29
30/*
31 * $Log: eql.c,v $
32 * Revision 1.2  1996/04/11 17:51:52  guru
33 * Added one-line eql_remove_slave patch.
34 *
35 * Revision 1.1  1996/04/11 17:44:17  guru
36 * Initial revision
37 *
38 * Revision 3.13  1996/01/21  15:17:18  alan
39 * tx_queue_len changes.
40 * reformatted.
41 *
42 * Revision 3.12  1995/03/22  21:07:51  anarchy
43 * Added capable() checks on configuration.
44 * Moved header file.
45 *
46 * Revision 3.11  1995/01/19  23:14:31  guru
47 * 		      slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
48 * 			(priority_Bps) + bytes_queued * 8;
49 *
50 * Revision 3.10  1995/01/19  23:07:53  guru
51 * back to
52 * 		      slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
53 * 			(priority_Bps) + bytes_queued;
54 *
55 * Revision 3.9  1995/01/19  22:38:20  guru
56 * 		      slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
57 * 			(priority_Bps) + bytes_queued * 4;
58 *
59 * Revision 3.8  1995/01/19  22:30:55  guru
60 *       slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
61 * 			(priority_Bps) + bytes_queued * 2;
62 *
63 * Revision 3.7  1995/01/19  21:52:35  guru
64 * printk's trimmed out.
65 *
66 * Revision 3.6  1995/01/19  21:49:56  guru
67 * This is working pretty well. I gained 1 K/s in speed.. now it's just
68 * robustness and printk's to be diked out.
69 *
70 * Revision 3.5  1995/01/18  22:29:59  guru
71 * still crashes the kernel when the lock_wait thing is woken up.
72 *
73 * Revision 3.4  1995/01/18  21:59:47  guru
74 * Broken set-bit locking snapshot
75 *
76 * Revision 3.3  1995/01/17  22:09:18  guru
77 * infinite sleep in a lock somewhere..
78 *
79 * Revision 3.2  1995/01/15  16:46:06  guru
80 * Log trimmed of non-pertinent 1.x branch messages
81 *
82 * Revision 3.1  1995/01/15  14:41:45  guru
83 * New Scheduler and timer stuff...
84 *
85 * Revision 1.15  1995/01/15  14:29:02  guru
86 * Will make 1.14 (now 1.15) the 3.0 branch, and the 1.12 the 2.0 branch, the one
87 * with the dumber scheduler
88 *
89 * Revision 1.14  1995/01/15  02:37:08  guru
90 * shock.. the kept-new-versions could have zonked working
91 * stuff.. shudder
92 *
93 * Revision 1.13  1995/01/15  02:36:31  guru
94 * big changes
95 *
96 * 	scheduler was torn out and replaced with something smarter
97 *
98 * 	global names not prefixed with eql_ were renamed to protect
99 * 	against namespace collisions
100 *
101 * 	a few more abstract interfaces were added to facilitate any
102 * 	potential change of datastructure.  the driver is still using
103 * 	a linked list of slaves.  going to a heap would be a bit of
104 * 	an overkill.
105 *
106 * 	this compiles fine with no warnings.
107 *
108 * 	the locking mechanism and timer stuff must be written however,
109 * 	this version will not work otherwise
110 *
111 * Sorry, I had to rewrite most of this for 2.5.x -DaveM
112 */
113
114#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
115
116#include <linux/capability.h>
117#include <linux/module.h>
118#include <linux/kernel.h>
119#include <linux/init.h>
120#include <linux/slab.h>
121#include <linux/timer.h>
122#include <linux/netdevice.h>
123#include <net/net_namespace.h>
124
125#include <linux/if.h>
126#include <linux/if_arp.h>
127#include <linux/if_eql.h>
128#include <linux/pkt_sched.h>
129
130#include <asm/uaccess.h>
131
132static int eql_open(struct net_device *dev);
133static int eql_close(struct net_device *dev);
134static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd);
135static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev);
136
137#define eql_is_slave(dev)	((dev->flags & IFF_SLAVE) == IFF_SLAVE)
138#define eql_is_master(dev)	((dev->flags & IFF_MASTER) == IFF_MASTER)
139
140static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave);
141
142static void eql_timer(unsigned long param)
143{
144	equalizer_t *eql = (equalizer_t *) param;
145	struct list_head *this, *tmp, *head;
146
147	spin_lock(&eql->queue.lock);
148	head = &eql->queue.all_slaves;
149	list_for_each_safe(this, tmp, head) {
150		slave_t *slave = list_entry(this, slave_t, list);
151
152		if ((slave->dev->flags & IFF_UP) == IFF_UP) {
153			slave->bytes_queued -= slave->priority_Bps;
154			if (slave->bytes_queued < 0)
155				slave->bytes_queued = 0;
156		} else {
157			eql_kill_one_slave(&eql->queue, slave);
158		}
159
160	}
161	spin_unlock(&eql->queue.lock);
162
163	eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
164	add_timer(&eql->timer);
165}
166
167static const char version[] __initconst =
168	"Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)";
169
170static const struct net_device_ops eql_netdev_ops = {
171	.ndo_open	= eql_open,
172	.ndo_stop	= eql_close,
173	.ndo_do_ioctl	= eql_ioctl,
174	.ndo_start_xmit	= eql_slave_xmit,
175};
176
177static void __init eql_setup(struct net_device *dev)
178{
179	equalizer_t *eql = netdev_priv(dev);
180
181	init_timer(&eql->timer);
182	eql->timer.data     	= (unsigned long) eql;
183	eql->timer.expires  	= jiffies + EQL_DEFAULT_RESCHED_IVAL;
184	eql->timer.function 	= eql_timer;
185
186	spin_lock_init(&eql->queue.lock);
187	INIT_LIST_HEAD(&eql->queue.all_slaves);
188	eql->queue.master_dev	= dev;
189
190	dev->netdev_ops		= &eql_netdev_ops;
191
192	/*
193	 *	Now we undo some of the things that eth_setup does
194	 * 	that we don't like
195	 */
196
197	dev->mtu        	= EQL_DEFAULT_MTU;	/* set to 576 in if_eql.h */
198	dev->flags      	= IFF_MASTER;
199
200	dev->type       	= ARPHRD_SLIP;
201	dev->tx_queue_len 	= 5;		/* Hands them off fast */
202	dev->priv_flags	       &= ~IFF_XMIT_DST_RELEASE;
203}
204
205static int eql_open(struct net_device *dev)
206{
207	equalizer_t *eql = netdev_priv(dev);
208
209	/* XXX We should force this off automatically for the user. */
210	netdev_info(dev,
211		    "remember to turn off Van-Jacobson compression on your slave devices\n");
212
213	BUG_ON(!list_empty(&eql->queue.all_slaves));
214
215	eql->min_slaves = 1;
216	eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */
217
218	add_timer(&eql->timer);
219
220	return 0;
221}
222
223static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave)
224{
225	list_del(&slave->list);
226	queue->num_slaves--;
227	slave->dev->flags &= ~IFF_SLAVE;
228	dev_put(slave->dev);
229	kfree(slave);
230}
231
232static void eql_kill_slave_queue(slave_queue_t *queue)
233{
234	struct list_head *head, *tmp, *this;
235
236	spin_lock_bh(&queue->lock);
237
238	head = &queue->all_slaves;
239	list_for_each_safe(this, tmp, head) {
240		slave_t *s = list_entry(this, slave_t, list);
241
242		eql_kill_one_slave(queue, s);
243	}
244
245	spin_unlock_bh(&queue->lock);
246}
247
248static int eql_close(struct net_device *dev)
249{
250	equalizer_t *eql = netdev_priv(dev);
251
252	/*
253	 *	The timer has to be stopped first before we start hacking away
254	 *	at the data structure it scans every so often...
255	 */
256
257	del_timer_sync(&eql->timer);
258
259	eql_kill_slave_queue(&eql->queue);
260
261	return 0;
262}
263
264static int eql_enslave(struct net_device *dev,  slaving_request_t __user *srq);
265static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq);
266
267static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
268static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
269
270static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc);
271static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc);
272
273static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
274{
275	if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG &&
276	    !capable(CAP_NET_ADMIN))
277	  	return -EPERM;
278
279	switch (cmd) {
280		case EQL_ENSLAVE:
281			return eql_enslave(dev, ifr->ifr_data);
282		case EQL_EMANCIPATE:
283			return eql_emancipate(dev, ifr->ifr_data);
284		case EQL_GETSLAVECFG:
285			return eql_g_slave_cfg(dev, ifr->ifr_data);
286		case EQL_SETSLAVECFG:
287			return eql_s_slave_cfg(dev, ifr->ifr_data);
288		case EQL_GETMASTRCFG:
289			return eql_g_master_cfg(dev, ifr->ifr_data);
290		case EQL_SETMASTRCFG:
291			return eql_s_master_cfg(dev, ifr->ifr_data);
292		default:
293			return -EOPNOTSUPP;
294	}
295}
296
297/* queue->lock must be held */
298static slave_t *__eql_schedule_slaves(slave_queue_t *queue)
299{
300	unsigned long best_load = ~0UL;
301	struct list_head *this, *tmp, *head;
302	slave_t *best_slave;
303
304	best_slave = NULL;
305
306	/* Make a pass to set the best slave. */
307	head = &queue->all_slaves;
308	list_for_each_safe(this, tmp, head) {
309		slave_t *slave = list_entry(this, slave_t, list);
310		unsigned long slave_load, bytes_queued, priority_Bps;
311
312		/* Go through the slave list once, updating best_slave
313		 * whenever a new best_load is found.
314		 */
315		bytes_queued = slave->bytes_queued;
316		priority_Bps = slave->priority_Bps;
317		if ((slave->dev->flags & IFF_UP) == IFF_UP) {
318			slave_load = (~0UL - (~0UL / 2)) -
319				(priority_Bps) + bytes_queued * 8;
320
321			if (slave_load < best_load) {
322				best_load = slave_load;
323				best_slave = slave;
324			}
325		} else {
326			/* We found a dead slave, kill it. */
327			eql_kill_one_slave(queue, slave);
328		}
329	}
330	return best_slave;
331}
332
333static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev)
334{
335	equalizer_t *eql = netdev_priv(dev);
336	slave_t *slave;
337
338	spin_lock(&eql->queue.lock);
339
340	slave = __eql_schedule_slaves(&eql->queue);
341	if (slave) {
342		struct net_device *slave_dev = slave->dev;
343
344		skb->dev = slave_dev;
345		skb->priority = TC_PRIO_FILLER;
346		slave->bytes_queued += skb->len;
347		dev_queue_xmit(skb);
348		dev->stats.tx_packets++;
349	} else {
350		dev->stats.tx_dropped++;
351		dev_kfree_skb(skb);
352	}
353
354	spin_unlock(&eql->queue.lock);
355
356	return NETDEV_TX_OK;
357}
358
359/*
360 *	Private ioctl functions
361 */
362
363/* queue->lock must be held */
364static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev)
365{
366	struct list_head *this, *head;
367
368	head = &queue->all_slaves;
369	list_for_each(this, head) {
370		slave_t *slave = list_entry(this, slave_t, list);
371
372		if (slave->dev == dev)
373			return slave;
374	}
375
376	return NULL;
377}
378
379static inline int eql_is_full(slave_queue_t *queue)
380{
381	equalizer_t *eql = netdev_priv(queue->master_dev);
382
383	if (queue->num_slaves >= eql->max_slaves)
384		return 1;
385	return 0;
386}
387
388/* queue->lock must be held */
389static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave)
390{
391	if (!eql_is_full(queue)) {
392		slave_t *duplicate_slave = NULL;
393
394		duplicate_slave = __eql_find_slave_dev(queue, slave->dev);
395		if (duplicate_slave)
396			eql_kill_one_slave(queue, duplicate_slave);
397
398		list_add(&slave->list, &queue->all_slaves);
399		queue->num_slaves++;
400		slave->dev->flags |= IFF_SLAVE;
401
402		return 0;
403	}
404
405	return -ENOSPC;
406}
407
408static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp)
409{
410	struct net_device *slave_dev;
411	slaving_request_t srq;
412
413	if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
414		return -EFAULT;
415
416	slave_dev  = dev_get_by_name(&init_net, srq.slave_name);
417	if (slave_dev) {
418		if ((master_dev->flags & IFF_UP) == IFF_UP) {
419			/* slave is not a master & not already a slave: */
420			if (!eql_is_master(slave_dev) &&
421			    !eql_is_slave(slave_dev)) {
422				slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL);
423				equalizer_t *eql = netdev_priv(master_dev);
424				int ret;
425
426				if (!s) {
427					dev_put(slave_dev);
428					return -ENOMEM;
429				}
430
431				memset(s, 0, sizeof(*s));
432				s->dev = slave_dev;
433				s->priority = srq.priority;
434				s->priority_bps = srq.priority;
435				s->priority_Bps = srq.priority / 8;
436
437				spin_lock_bh(&eql->queue.lock);
438				ret = __eql_insert_slave(&eql->queue, s);
439				if (ret) {
440					dev_put(slave_dev);
441					kfree(s);
442				}
443				spin_unlock_bh(&eql->queue.lock);
444
445				return ret;
446			}
447		}
448		dev_put(slave_dev);
449	}
450
451	return -EINVAL;
452}
453
454static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp)
455{
456	equalizer_t *eql = netdev_priv(master_dev);
457	struct net_device *slave_dev;
458	slaving_request_t srq;
459	int ret;
460
461	if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
462		return -EFAULT;
463
464	slave_dev = dev_get_by_name(&init_net, srq.slave_name);
465	ret = -EINVAL;
466	if (slave_dev) {
467		spin_lock_bh(&eql->queue.lock);
468
469		if (eql_is_slave(slave_dev)) {
470			slave_t *slave = __eql_find_slave_dev(&eql->queue,
471							      slave_dev);
472
473			if (slave) {
474				eql_kill_one_slave(&eql->queue, slave);
475				ret = 0;
476			}
477		}
478		dev_put(slave_dev);
479
480		spin_unlock_bh(&eql->queue.lock);
481	}
482
483	return ret;
484}
485
486static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
487{
488	equalizer_t *eql = netdev_priv(dev);
489	slave_t *slave;
490	struct net_device *slave_dev;
491	slave_config_t sc;
492	int ret;
493
494	if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
495		return -EFAULT;
496
497	slave_dev = dev_get_by_name(&init_net, sc.slave_name);
498	if (!slave_dev)
499		return -ENODEV;
500
501	ret = -EINVAL;
502
503	spin_lock_bh(&eql->queue.lock);
504	if (eql_is_slave(slave_dev)) {
505		slave = __eql_find_slave_dev(&eql->queue, slave_dev);
506		if (slave) {
507			sc.priority = slave->priority;
508			ret = 0;
509		}
510	}
511	spin_unlock_bh(&eql->queue.lock);
512
513	dev_put(slave_dev);
514
515	if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t)))
516		ret = -EFAULT;
517
518	return ret;
519}
520
521static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
522{
523	slave_t *slave;
524	equalizer_t *eql;
525	struct net_device *slave_dev;
526	slave_config_t sc;
527	int ret;
528
529	if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
530		return -EFAULT;
531
532	slave_dev = dev_get_by_name(&init_net, sc.slave_name);
533	if (!slave_dev)
534		return -ENODEV;
535
536	ret = -EINVAL;
537
538	eql = netdev_priv(dev);
539	spin_lock_bh(&eql->queue.lock);
540	if (eql_is_slave(slave_dev)) {
541		slave = __eql_find_slave_dev(&eql->queue, slave_dev);
542		if (slave) {
543			slave->priority = sc.priority;
544			slave->priority_bps = sc.priority;
545			slave->priority_Bps = sc.priority / 8;
546			ret = 0;
547		}
548	}
549	spin_unlock_bh(&eql->queue.lock);
550
551	dev_put(slave_dev);
552
553	return ret;
554}
555
556static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp)
557{
558	equalizer_t *eql;
559	master_config_t mc;
560
561	memset(&mc, 0, sizeof(master_config_t));
562
563	if (eql_is_master(dev)) {
564		eql = netdev_priv(dev);
565		mc.max_slaves = eql->max_slaves;
566		mc.min_slaves = eql->min_slaves;
567		if (copy_to_user(mcp, &mc, sizeof (master_config_t)))
568			return -EFAULT;
569		return 0;
570	}
571	return -EINVAL;
572}
573
574static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp)
575{
576	equalizer_t *eql;
577	master_config_t mc;
578
579	if (copy_from_user(&mc, mcp, sizeof (master_config_t)))
580		return -EFAULT;
581
582	if (eql_is_master(dev)) {
583		eql = netdev_priv(dev);
584		eql->max_slaves = mc.max_slaves;
585		eql->min_slaves = mc.min_slaves;
586		return 0;
587	}
588	return -EINVAL;
589}
590
591static struct net_device *dev_eql;
592
593static int __init eql_init_module(void)
594{
595	int err;
596
597	pr_info("%s\n", version);
598
599	dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", eql_setup);
600	if (!dev_eql)
601		return -ENOMEM;
602
603	err = register_netdev(dev_eql);
604	if (err)
605		free_netdev(dev_eql);
606	return err;
607}
608
609static void __exit eql_cleanup_module(void)
610{
611	unregister_netdev(dev_eql);
612	free_netdev(dev_eql);
613}
614
615module_init(eql_init_module);
616module_exit(eql_cleanup_module);
617MODULE_LICENSE("GPL");
618